100-Class Animal Bounding Box Annotation Dataset (A100-Det)

This dataset constructs a rectangular box labeled dataset containing 100 animal species, with a sample size of 100-130 for each animal species, achieving wide coverage of animal species and sufficient supply of single class samples. This dataset has a large scope and covers a large number of animals, so a more specific classification of animal species has not been made more strictly (such as treating great white sharks, hammerhead sharks, grey eyed sharks, etc. as sharks). If more detailed classification is needed for research, please supplement the dataset yourself. As of now, there is no publicly available dataset that can simultaneously meet the core requirements of "100 or more animals, single class sample size ≥ 100, high-precision rectangular box annotation". This dataset fills this research gap and can provide high-quality training and testing data support for the development and optimization of animal target detection models, especially for the promotion of research in niche animal detection, multi animal mixed scene detection, and other subdivision directions. For more information on the dataset, please refer to the README.md in the dataset compression package Data collection and cleaning: All images are searched from Bing images（ https://cn.bing.com/images ）Obtain and strictly follow copyright regulations. The specific process is as follows: Visit the Bing image search page, use the filter to select the "Free to modify, share, and use" permission, and ensure that the image usage is compliant; Taking Antelope as an example, the search link is: https://cn.bing.com/images/search?q=antelope&qs=n&form=QBIR&qft=%20filterui%3Alicense -L2_L3_L5_L6， Search for pictures of 100 animals in this order; Use the Microsoft Edge browser extension ImageAssistant to batch extract images from the current search page and download them; Manual cleaning: Remove blurry, heavily occluded, low resolution, unclear animal subjects, or images with copyright disputes to ensure the usability of each image. Annotation: Adopting a semi-automatic annotation method of "AI assisted+manual correction", balancing annotation efficiency and accuracy. The specific process is as follows: Labeling tool: Use the open-source labeling tool X-AnyLabeling（ https://github.com/CVHub520/X-AnyLabeling ）； AI assisted annotation: Call the YOLOv8m model, set the confidence threshold to 0.4 and the Intersection over Union (IoU) threshold to 0.8, and perform batch pre annotation on all cleaned images; Manual correction: Check the AI pre annotation results one by one, and manually correct any deviations, omissions, or errors in the annotation box positions; Manually annotate animal targets that are not detected by AI to ensure that the animal subjects in each image are accurately labeled; Format export: After the annotation is completed, keep the original JSON annotation file of X-AnyLabeling and export it as a VOC XML format file to meet the training requirements of mainstream object detection models such as YOLO series, Faster R-CNN, etc